The classical central paradigm of structural biology links a protein’s sequence to its structure and function but overlooks conformational fluctuation that is integral to protein function. We propose a graph neural network model based on gated attention that explicitly incorporates protein dynamics via physics-based models to predict protein crystallization propensity. We compare results to all-atom molecular dynamics simulations of flexible, disordered human tropoelastin and ordered, globular human lysyl oxidase-like protein. Our findings show that fluctuating residues correlate with locally maximal attention scores in the neural network. By methodically truncating the sequences, we establish correlations between dynamical and physicochemical molecular properties and protein crystallization propensity. Accounting for comprehensive biological mechanisms, our tool can facilitate the rational design of protein sequences that lead to diffraction-quality crystals. Our study showcases the integration of physics-based and machine learning models for structure and property prediction, expanding the classical paradigm of structural biology.

结构生物学的经典中心范式将蛋白质的序列与其结构和功能联系起来，但忽略了蛋白质功能不可或缺的构象波动。我们提出了一种基于门控注意力的图神经网络模型，该模型通过基于物理的模型明确结合蛋白质动力学来预测蛋白质结晶倾向。我们将结果与灵活的、无序的人类原弹性蛋白和有序的、球状的人类赖氨酰氧化酶样蛋白的全原子分子动力学模拟进行比较。我们的研究结果表明，波动的残基与神经网络中的局部最大注意力分数相关。通过有条不紊地截断序列，我们建立了动力学和物理化学分子特性与蛋白质结晶倾向之间的相关性。考虑到全面的生物学机制，我们的工具可以促进蛋白质序列的合理设计，从而产生衍射质量的晶体。我们的研究展示了基于物理和机器学习模型的结构和性质预测的集成，扩展了结构生物学的经典范式。